To recognize images under various view-angle in noised background is one of the most important and challenging problems that need be resolved. An approach using affine invariant features in frequency domain for object recognition in low signal noise ratio (SNR) images is presented. The relationship of affine transforms between spatial domain and frequency domain is analyzed. That is, the effect of affine transformation on the spectrum is almost the same as the affine transform on the object in spatial domain except for two major differences: Firstly, the spectrum is inversely scaled and slanted; Secondly, shape translations parallel to the image plane do not affect the spectrum. In the preprocessing procedure, noise is reduced by convoluting the input image with a low-pass filter. Then the spectral signatures at low-to-median frequency are extracted using pseudo-log sampling, which is useful to reduce the influences of both view-point and illumination variation and to further suppress high frequency noises. A neural network is trained with these features to extract affine invariant features and to recognize objects at different poses. Comparing the proposed method with other two Gabor-based methods, experimental results show that more than 90% images can be recognized correctly even when SNR drops bellow -20dB and this approach is much better than the Gabor-based methods in low SNR images.